Streaming Scalable Videos over Multi-Hop Cognitive Radio Networks

Transcription

1 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 11, NOVEMBER Streaing Scaabe Videos over Muti-Hop Cognitive Radio Networks Dongin Hu and Shiwen Mao, Senior Meber, IEEE Abstract We investigate the probe of streaing utipe videos over uti-hop cognitive radio (CR) networks. Fine-Granuarity-Scaabiity (FGS) and Mediu-Grain-Scaabe (MGS) videos are adopted to accoodate the heterogeneity aong channe avaiabiities and dynaic network conditions. We obtain a ixed integer noninear prograing (MINLP) probe foruation, with objectives to axiize the overa received video quaity and to achieve fairness aong the video sessions, whie bounding the coision rate with priary users under the presence of spectru sensing errors. We first sove the MINLP probe using a centraized sequentia fixing agorith, and derive upper and ower bounds for the objective vaue. We then appy dua decoposition to deveop a distributed agorith and prove its optiaity and convergence conditions. The proposed agoriths are evauated with siuations and are shown to be effective in supporting concurrent scaabe video sessions in uti-hop CR networks. Index Ters Cross-ayer optiization, dynaic spectru access, distributed agorith, uti-hop cognitive radio networks, video streaing. I. INTRODUCTION A cognitive radio (CR) is an advanced radio device with interface(s) to sense the radio environent, an inteigent agent for decision-aking based on radio environent and past experience, and a frequency-agie radio that can be tuned to a wide range of frequency bands and operate fro there. CR represents a paradig change in spectru reguation and access, fro excusive use by icensed, or priary, users to sharing spectru with, and dynaic spectru access for unicensed, or secondary, users. It has profound ipact on how future wireess networks wi be designed and operated. The high potentia of CRs has attracted substantia interest. The ainstrea CR research has focused on deveoping effective spectru sensing and access techniques (see [1] and [2] and reference therein). Athough considerabe understandings have been gained on various aspects of CR, the iportant probe of guaranteeing appication perforance has not been the focus of ajor CR research. To this end, we find spectru-intensive and rate-adaptive utiedia, or video as Manuscript received January 23, 2010; revised June 1, 2010 and August 30, 2010; accepted August 31, The associate editor coordinating the review of this paper and approving it for pubication was Dr. D. Tarchi. This work was supported in part by the US Nationa Science Foundation (NSF) under Grants CNS , IIP , and ECCS , and through the Wireess Internet Center for Advanced Technoogy (WICAT) at Auburn University. The authors are with the Departent of Eectrica and Coputer Engineering, Auburn University, Auburn, AL (e-ai: dzh0003@auburn.edu, sao@ieee.org). Coor versions of one or ore of the figures in this paper are avaiabe onine at Digita Object Identifier /TWC Fig /10$25.00 c 2010 IEEE Priary network base station Priary network user Secondary network user Iustration of the uti-hop video CR network architecture. a reference appication, akes exceent use of the enhanced spectru efficiency in CR networks. Unike data, where each bit shoud be deivered, video is oss toerant and rate adaptive. They are highy suited for CR networks, where the avaiabe bandwidth heaviy depends on priary user behavior. Gracefu degradation of video quaity can be achieved as spectru opportunities change over tie. In our prior work [3], we investigated the probe of uticasting Fine-Granuarity-Scaabity (FGS) video in an infrastructure-based CR network and deonstrated the feasibiity of video over CR networks. In this paper, we study the ore chaenging probe of video over uti-hop CR networks. As iustrated in Fig. 1, we consider an infrastructureess CR network co-ocated with one or ore fixed priary networks. CR users non-intrusivey expoit white spaces in the icensed bands for streaing utipe videos. The objective is two-fod: to axiize the overa video quaity and to achieve fairness aong the concurrent video sessions, subject to bounded interference to priary users. We adopt FGS videos to accoodate heterogeneous channe avaiabiities and dynaic network conditions [4]. FGS video is coded into a base ayer (BL) and an enhanceent ayer (EL). The EL can be truncated at any bit ocation, whie a the reaining bits are sti usefu for decoding. This feature sipifies the design of video streaing systes. We aso consider H.264/SVC Mediu-Grain-Scaabe (MGS) videos in this paper. MGS is shown to achieve better rate-distortion perforance over MPEG-4 FGS, athough MGS ony has Network Abstraction Layer (NAL) unit-based granuarity [5]. In order to ode and guarantee end-to-end video perforance, we adopt the apify-and-forward reay approach for video data, which is we-studied in the context of cooperative counications [6]. Specificay, each CR node is equipped

2 3502 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 11, NOVEMBER 2010 with two transceivers operating on orthogona channes. During data transission, a reay CR node receives data fro its upstrea node using one transceiver on one channe, whie siutaneousy apifies and forwards the received data to its downstrea node using the other transceiver operating on a different channe. This is equivaent to estabishing a virtua tunne through a uti-hop uti-channe path, as iustrated in Fig. 2. It is aso anaogous to the cut-through switching approach for packet switching networks [7]. In addition to aowing a neat foruation of the chaenging uti-hop video streaing probe, this approach aso satisfies video s needs for ow atency, ow jitter, and high bandwidth. Its feasibiity and practica considerations for uti-hop wireess networks have been addressed in [8]. The target probe is non-trivia due to the additiona diension of network dynaics (i.e., channe avaiabiity) and the additiona uncertainty (i.e., spectru sensing and sensing errors) found in CR networks. The stringent QoS requireents necessitate cross-ayer optiization. The ack of centraized contro aso cas for distributed agoriths. We foruate streaing utipe videos over a uti-hop CR network as a ixed integer noninear prograing (MINLP) probe, considering iportant design factors such as spectru sensing and sensing errors, spectru access and priary user protection, video quaity and fairness, channe scheduing, and path seection. We first deveop a centraized sequentia fixing agorith to derive upper and ower bounds for the achievabe video quaity. These bounds provide usefu insights on perforance iits of the CR video syste. We then decopose the MINLP probe into a channe scheduing probe and a path seection probe. The channe scheduing probe is soved with a greedy agorith. For path seection, we appy dua decoposition and deveop a distributed agorith. We prove the optiaity of the proposed approach and derive the convergence condition for the distributed agorith. The agoriths are evauated with extensive siuations. The distributed agorith is shown to be highy effective for supporting concurrent video sessions in uti-hop CR networks, as it can achieve a perforance cose to that of the centraized agorith as we as the upper bound in the cases exained. The reainder of this paper is organized as foows. The syste ode is described in Section II. We present the probe foruation and deveop the centraized agorith in Section III. We derive the distributed agorith and anayze its optiaity and convergence perforance in Section IV. Siuation resuts are presented in Section V and reated work is discussed in Section VI. Section VII concudes the paper. II. SYSTEM MODEL A. Network Mode We consider a spectru band consisting of M orthogona channes with identica bandwidth [9]. As shown in Fig. 1, the channes are shared by K priary networks and one uti-hop CR network. The priary network base stations provide data or utiedia service to priary users. There is no fixed infrastructure in the CR network; secondary users non-intrusivey expore the spectru opportunities for unicast video counications. 1) Priary Networks: We assue that the M channes are aocated to K priary networks, which cover different service areas. We further assue that the priary systes use a synchronous sot structure as in prior work [2], [10]. Due to priary user transissions, the occupancy of each channe evoves foowing a discrete-tie Markov process, as vaidated by recent easureent studies [2], [10], [11]. In priary network k, the status of channe in tie sot t is denoted by S k (t) with ide (i.e., Sk (t) =0) and busy (i.e., S(t) k =1) states. Let λ k and μ k be the transition probabiity of reaining in state 0 and that fro state 1 to 0, respectivey, for channe in priarynetwork k. The utiization of channe in priary network k, denoted by η k =Pr(Sk =1),is η k = i 1 T T t=1 T Sk (t) = 1 λ k 1 λ k +. (1) μk 2) The Muti-hop CR Network: Consider a uti-hop CR network co-ocated with the priary networks, within which N CR nodes are streaing S rea-tie videos. Let U k denote the set of CR nodes that are ocated within the coverage of priary network k. A video session ay be reayed by utipe CR nodes if source z is not a one-hop neighbor of destination d. We assue a coon contro channe for the CR network [10]. We aso assue the tiescae of the priary channe process (or, the tie sot durations) is arger than the broadcast deays on the contro channe, such that feedbacks of channe inforation can be received at the source nodes in a tiey anner. For CR users, each tie sot consists of three phases: the spectru sensing phase, the data transission phase, and the acknowedgent phase. Assue that each CR user has two transceivers. In the sensing phase, one transceiver is used to sense one of the M channes, whie the other is tuned to the contro channe to exchange channe inforation with other CR users. Each video source coputes the optia path seection and channe scheduing based on sensing resuts. In the transission phase, the channes assigned to a video session at each ink aong the path for a virtua tunne connecting source z and destination d, as iustrated in Fig. 2. Each node can use one or ore than one channes to counicate with other nodes using a channe aggregation technique such as Orthogona Frequency Division Mutipexing (OFDM) [10], [12]. When utipe channes are avaiabe on a the inks aong a path, utipe tunnes can be estabished and used siutaneousy for a video session. In the acknowedgent phase, the destination sends ACK to the source for successfuy received video packets through the sae tunne. We adopt apify-and-forward for video transission [6]. During the transission phase, one transceiver of the reay node receives video data fro the upstrea node on one channe, whie the other transceiver of the reay node apifies and forwards the data to the downstrea node on a different, orthogona channe. There is no need to store video packets at the reay nodes. Error detection/correction wi be perfored at the destination node. As a resut, we can transit through the tunne a bock of video data with iniu deay and jitter in one tie sot. Copared to the traditiona hop-centric approach, this schee greaty reduces the coision, contention, processing,

3 HU and MAO: STREAMING SCALABLE VIDEOS OVER MULTI-HOP COGNITIVE RADIO NETWORKS 3503 tunne ide busy tunne busy Source z Interediate node Destination d Sensing Transission ACK Fig. 2. busy The cut-through switching ode for video data. and queueing deay induced at reay nodes [7], [8]. It is suitabe for rea-tie data with tight deay and jitter requireents. It is especiay aicabe for FGS video, since a corrupted packet ay sti be usefu for decoding. The viabiity, protocoreated issues, and practica considerations of this approach are addressed in [8]. The chaenging issue, however, is how to set up the tunnes, whie the avaiabe channes at the reays evove over tie due to priary user transissions. We wi address this issue in Section IV. B. Spectru Sensing Athough precise and tiey channe state inforation is highy desirabe, continuous fu-spectru sensing is hardware deanding. Without oss of generaity, we assue each CR user periodicay chooses one channe to sense in each tie sot [13]. The index of the chosen channe by user i in tie sot t is: M i (t) =[M i (0) + t 1] od M, wherem i (0) is the index of the channe sensed in tie sot 0. There are two types of spectru sensing errors: with a fase aar, a spectru opportunity wi be wasted, whie a iss detection ay ead to coision with priary users. Without oss of generaity, we assue that the sae spectru sensing echanis is used with identica sensing error probabiities for a CR users. Let ε and δ denote the probabiities of fase aar and iss detection on channe, respectivey. For tie sot t, wehave: P (W i (t) =1 S(t) k =0)=ε, =1,,M P (Wi (t) =0 S(t) k =1)=δ, =1,,M, where Wi (t) is user i s sensing resut for channe. In a uti-hop CR network, the sensing resuts fro various users ay be different. Denote H 0 as the hypothesis that channe in priary network k is ide, and H 1 the hypothesis that channe in priary network k is busy in tie sot t. The conditiona probabiity that channe is avaiabe in priary network k, denoted by a k (t), can be derived as in (3), where θi represents a specific sensing resut (0 or 1), U k is the subset of users in U k (i.e., the set of CR nodes that are ocated within the coverage of priary network k) that sense channe, u k is the nuber of users in U k observing channe is ide, π k represents the history of channe in priary (2) network k, 1 and φ k and φk are defined as: φ k = P (W i =0 H1) δ P (Wi =0 H0) = 1 ε, when θi =0 φ k = P (W i =1 H1) (4) 1 δ P (Wi =1 H0) = ε, when θi =1. The third equaity equaity in (3) is due to independent sensing processes. The fourth equaity is because sensing processes are independent of channe history. Based on the Markov chain channe ode, we have (5), which can be recursivey expanded: [ Pr(H0 π)=λ k k a k (t 1) + μ k 1 a k (t 1) ] Pr(H 1 π k )=1 Pr(H 0 π k ). (5) C. Spectru Access and Interference Modeing Based on spectru sensing resuts, a CR user deterines which channe(s) to access for transission of video data. Let κ k be a threshod for spectru access: channe is considered ide if the estiate a k is greater than the threshod, and is considered busy otherwise. The avaiabiity of channe in priary network k, denoted as A k,is A k 0, a = k κ k (6) 1, otherwise. For each channe, we can cacuate the probabiity of coision with priary users as: Pr(A k =0 H 1)= ( ) U k i ψ (1 δ k i ) U k i (δ ) i, (7) where set ψ k is defined as: [ ψ k = i 1+φ i φ U k i Pr(H 1 π) k ] 1 } Pr(H 0 π) k κ k. (8) For non-intrusive spectru access, the coision probabiity shoud be bounded with a prescribed threshod γ k. A higher spectru access threshod κ k wi reduce the potentia interference with priary users, but increase the chance of wasting transission opportunities. For a given coision toerance γ k, we can sove Pr(A k =0 H 1 )=γ k for κ k. The objective is to axiize CR users spectru access without exceeding the axiu coision probabiity with priary users. Let Ω i,j be the set of avaiabe channes at ink i, j}. Assuing i U k and j U k,wehave } Ω i,j = =0and Ak =0. (9) A k D. Link and Path Statistics Due to the apify-and-forward approach for video data transission, there is no queueing deay at interediate nodes. Assue each ink has a fixed deay ω i,j (i.e., processing and propagation deays). Let P A be the set of a possibe paths fro z to d. For a given deay requireent T th,thesetof feasibe paths P for video session can be deterined as: P = P } i,j} P ω i,j T th, P P A. (10) 1 π k represents the avaiabiity of channe in priary network k in the previous tie sot. If the channe was used in that tie sot, π k can be readiy deterined as 0 or 1, since the channe state was known (i.e., with or without ACKs). Otherwise, π k can be estiated in the for of ak (t 1) as in (3).

4 3504 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 11, NOVEMBER 2010 a k (t) =Pr(H 0 Wi = θi,i U,π k )= k Pr(Wi = θi,i Uk H 0,π k )Pr(H 0 π k ) s 0,1} Pr(W i = θi,i Uk H s,π k )Pr(H s π k ) Pr(H 0 π k = ) i U Pr(W k i = θi H 0,π k ) s 0,1} Pr(H s π k ) i U Pr(W k i = θi H s,π k ) = Pr(H 0 πk ) i U Pr(W k i = θi H 0) s 0,1} Pr(H s π k ) i U Pr(W k i = θi H s) 1 = 1+ Pr(H 1 π) k Pr(Wi = θi H [ 1) Pr(H 0 π) k Pr(Wi = θi H = 1+ ( φ k ) u k ( φ k ) U k uk Pr(H 1 π k ] 1 ) 0) Pr(H 0 π) k. (3) i U k Let p i,j be the packet oss rate on channe at ink i, j}. A packet is successfuy deivered over ink i, j} if there is no oss on a the channes that were used for transitting the packet. The ink oss probabiity p i,j can be derived as: p i,j =1 M (1 p i,j )I, (11) where M is set of icensed channes and I is an indicator: I =1if channe is used for the transission, and I =0 otherwise. Assuing independent ink osses, the end-to-end oss probabiity for path P h P can be estiated as: p h =1 i,j} P h E. Video Perforance Measure (1 p i,j ). (12) As discussed, both FGS and MGS videos are highy suitabe for dynaic CR networks [3]. With FGS or MGS coding, each video is encoded into one base ayer with rate R b and one enhanceent ayer with rate R e. The tota bit rate for video is R = R b + Re. We consider peak-signa-noise-ratio (PSNR) (in db) of reconstructed videos. As in prior work [3], [14], the average PSNR of video, denoted as Q, can be estiated as: Q (R )=Q b + β (R R b )=Q0 + β R, (13) where Q b is the resuting PSNR when the base ayer is decoded aone, β a constant depending on the video sequence and codec setting, and Q 0 = Qb β R b.weverified the ode (13) with severa test video sequences using the MPEG-4 FGS codec and the H.264/SVC MGS codec and found it is highy accurate. The resuts are oitted for brevity. Due to the rea-tie nature, we assue that each group of pictures (GOP) ust be deivered during the next GOP window, which consists of N G tie sots. Beyond that, overdue data fro the current GOP wi be useess and wi be discarded. We further assue fixed packet ength L p. Therefore the average rate in a GOP window is R =(N p L p )/(N G T s ), where T s is the tie sot duration and N p is the nuber of successfuy deivered packets. We aso ai to achieve fairness aong the concurrent video sessions. It has been shown that proportiona fairness can be achieved by axiizing the su of ogariths of video PSNRs (i.e., utiities). Therefore, our objective is to axiize the overa syste utiity, i.e., axiize: U (R )= og(q (R )). (14) x,h,r i,j, = III. PROBLEM STATEMENT A. Muti-hop CR Network Video Streaing Probe For the syste described in Section II, the probe of video over uti-hop CR networks consists of path seection for each video session and channe scheduing for each CR node aong the chosen paths. We define two sets of index variabes. For channe scheduing, we have 1, at ink i, j}, if channe is assigned to tunne r in path P h 0, otherwise. (15) For path seection, we have y h 1, if video session seects path P h = P (16) 0, otherwise, Note that the indicators, x,h,r i,j, and yh, are not independent. If y h =0for path P h, a the x,h,r i,j, s on that path are 0. If ink i, j} is not on path P h, a its x,h,r i,j, s are aso 0. For ink i, j} on path P h, we can ony choose those avaiabe channes in set Ω i,j to schedue video transission. That is, we have x,h,r i,j, 0, 1} if Ω i,j, andx,h,r i,j, =0otherwise. In the rest of the paper, we use x and y to represent the vector fors of x,h,r i,j, and yh, respectivey. As discussed, the objective is to axiize the expected utiity su at the end of N G tie sots, as given in (14). Since og(q (E[R (0)])) is a constant, (14) is equivaent to the su of utiity increents of a the tie sots, as og(q (E[R (N G )])) og(q (E[R (0)])) = t og(q (E[R (t)])) og(q (E[R (t 1)]))}.(17) Therefore, (14) wi be axiized if we axiize the expected utiity increent during each tie sot, which can be written as: og(q (E[R (t)])) og(q (E[R (t 1)])) = ( ) og E[R (t)] E[R (t 1)] 1+β Q (E[R (t 1)]) = y h og 1+ β L p x,h,r z,z, N h P r G T s Q t 1 (1 p r,h ) = ( ) y h og 1+ρ t x,h,r z,z,(1 pr,h ), h P r where z is the next hop fro z on path P h, pr,h is the packet oss rate on tunne r of path P h, Qt 1 = Q (E[R (t 1)]), and ρ t = β L p /(N G T s Q t 1 ).

5 HU and MAO: STREAMING SCALABLE VIDEOS OVER MULTI-HOP COGNITIVE RADIO NETWORKS 3505 Fro (11) and (12), the end-to-end packet oss rate for tunne r on path P h is: p r,h =1 i,j} P M (1 h p i,j) x,h,r i,j,. (18) We assue that each tunne can ony incude one channe on each ink. When there are utipe channes avaiabe at each ink aong the path, a CR source node can set up utipe tunnes to expoit the additiona bandwidth. We then have the foowing constraint: x,h,r i,j, 1, i, j} Ph. (19) Considering avaiabiity of the channes, we further have, r x,h,r i,j, Ω i,j, i, j} P h, (20) where Ω i,j is the nuber of avaiabe channes on ink i, j} defined in (9). As discussed, each node is equipped with two transceivers: one for receiving and the other for transitting video data during the transission phase. Hence a channe cannot be used to receive and transit data siutaneousy at a reay node. We have for each channe : r x,h,r i,j, + r x,h,r j,k, 1,,, h P, i, j}, j, k} P h. (21) Let n h be the nuber of tunnes on path P h. For each source z and each destination d, the nuber of schedued channes is equa to n h. We have for each source node r x,h,r z,z, = nh y h, h P,. (22) Let d be the ast hop to destination d on path P h,wehave for each destination node r x,h,r d,d, = nh yh, h P,. (23) At a reay node, the nuber of channes used to receive data is equa to that of channes used to transit data, due to fow conservation and apify-and-forward. At reay node j for session, assue i, j} P h and j, k} P h.wehave, r x,h,r i,j, = r x,h,r j,k,, h P,, i, j}, j, k} P h. (24) We aso consider hardware-reated constraints on path seection. We suarize such constraints in the foowing genera for for ease of presentation: h P w g,h yh 1, g. (25) To sipify exposition, we choose at ost one path in P for video session. Such a singe path routing constraint can be expressed as h yh 1, which is a specia case of (25) where w,h 1 =1for a h, andwg,h =0for a g = 1, =, andh. We can aso have h yh ξ to aow up to ξ paths for each video session. In order to achieve optiaity in the genera case of uti-path routing, an optia scheduing agorith shoud be designed to dispatch packets to paths with different conditions (e.g., different nuber of tunnes and deays). There are aso disjointedness constraints for the chosen paths. This is because each CR node is equipped with two transceivers and both wi be used for a video session if it is incuded in a chosen path. Such disjointedness constraint is aso a specia case of (25) with the foowing definition for w g,h for each CR node g: w g 1, if node g path P,h = h 0, otherwise, (26) Finay we foruate the probe of uti-hop CR network video streaing (OPT-CRV) as: ( ) ax: y h og 1+ρ t x,h,r z,z,(1 pr,h) (27) h P r subject to: (15) (25). B. Centraized Agorith and Upper/Lower Bounds Probe OPT-CRV is in the for of MINLP (without continuous variabes), which is NP-hard in genera. We first describe a centraized agorith to derive perforance bounds in this section, and then present a distributed agorith based on dua decoposition in the next section. We first obtain a reaxed non-inear prograing (NLP) are are treated as nonnegative rea nubers. It can be shown that the reaxed probe has a concave object function and the constraints are convex. This reaxed probe can be soved using a constrained noninear optiization probe sover. If a the variabes are integer in the soution, then we have the exact optia soution. Otherwise, we obtain an infeasibe soution, which produces an upper bound for the probe. This is given in Lines 1 2 intabei. We aso deveop a sequentia fixing agorith (SF) for soving OPT-CRV. The pseudo-code is given in Tabe I. SF iterativey soves the reaxed probe, fixing one or ore integer variabes after each iteration [3], [15]. In Tabe I, Lines version of OPT-CRV. The binary variabes x,h,r i,j, and yh reaxed to take vaues in [0,1]. The integer variabes n h 3 7 fix the path seection variabes y h, and Lines 8 17 fix the channe scheduing variabes x,h,r i,j, and tunne variabes n h. The tunne variabes nh can be coputed using (22) after x,h,r i,j, and yh are soved. When the agorith terinates, it produces a feasibe soution that yieds a ower bound for the objective vaue. IV. DUAL DECOMPOSITION SF is a centraized agorith requiring goba inforation. It ay not be suitabe for uti-hop wireess networks, athough the upper and ower bounds provide usefu insights on the perforance iits. In this section, we deveop a distributed agorith for Probe OPT-CRV and anayze its optiaity and convergence perforance. A. Decopose Probe OPT-CRV Since the doains of x,h,r i,j, defined in (19) (24) for different paths do not intersect with each other, we can decopose Probe OPT-CRV into two subprobes. The first subprobe deas with channe scheduing for axiizing the

6 3506 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 11, NOVEMBER 2010 TABLE I SEQUENTIAL FIXING ALGORITHM (SF) FOR PROBLEM OPT-CRV 1 : Reax integer variabes x,h,r i,j,, yh,andnh ; 2 : Sove the reaxed probe using a constrained NLP sover; 3: if (there is y h not fixed) 4 : Find the argest y h,where[,h ] = arg axy h}, and fixitto1; 5: Fixothery h s according to constraint (25); 6: GotoStep2; 7: end if 8: if (there is x,h,r i,j, not fixed) 9 : Find the argest x,h,r i,j,,where[i,j,,,h,r ]= arg axx,h,r i,j, },andsetitto1; 10: Fix other x,h,r i,j, s according to the constraints; 11: if (there is other variabe that is not fixed) 12: Go to Step 2; 13: ese 14: Fix n h s based on x and y; 15: Exit with feasibe soution x, y, n}; 16: end if 17: end if expected utiity on a chosen path P h.wehavethechanne scheduing probe (OPT-CS) as: H h =ax x r x,h,r z,z,(1 pr,h) (28) subject to: (19) (24), x,h,r z,z, 0, 1}, for a, h, r,. In the second part, optia paths are seected to axiize the overa objective function. Letting F h =og ( ) 1+ρ T Hh,we have the foowing path seection probe (OPT-PS): axiize: subject to: f(y) = h F h yh (29) h P w g,h yh y h 0, 1}, for a, h. 1, for a g TABLE II GREEDY ALGORITHM FOR CHANNEL SCHEDULING 1 : Initiaization: tunne r =1, ink i, j} s fro z to d ; 2: if ( Λ i,j == 0) 3 : Exit; 4: ese if ( Λ i,j == 1) 5 : Assign the singe channe in Λ i,j,, to tunne r; 6 Check neighboring ink k, i}; 7: if (p k,i Λ k,i) 8 : Reove p k,i fro Λ k,i, i k, j i andgotostep2; 9: ese 10: Go to Step 13; 11: end if 12: ese 13: Put Λ i,j in set Λ h ; 14: if (node j is not destination d ) 15: i j, j v; 16: Go to Step 2; 17: end if 18: end if 19: whie (Λ h is not epty) 20: Find the axiu vaue p i,j in set Λ h i,j, } =arginp i,j }; 21: Assign channe to tunne r; 22: Reove set Λ i,j fro set Λh ; 23: Check neighboring ink k, i} and j, v}; 24: if (p k,i Λ k,i and Λ k,i Λ h ) 25: Reove p k,i fro Λ k,i; 26: if (Λ k,i is epty) 27: Exit; 28: end if 29: end if 30: if (p j,v Λ j,v and Λ j,v Λ h ) 31: Reove p j,v fro Λ j,v; 32: if (Λ j,v is epty) 33: Exit; 34: end if 35: end if 36: end whie 37: Copute the next tunne: r r +1andgotoStep2; B. Sove the Channe Scheduing Subprobe We have the foowing resut for assigning avaiabe channes at a reay node. Theore 1: Consider three consecutive nodes aong a path, denoted as nodes i, j, and k. Ide channes 1 and 2 are avaiabe at ink i, j} and ide channes 3 and 4 are avaiabe at ink j, k}. Assue the packet oss rates of the four channes satisfy p 1 i,j >p2 i,j and p3 j,k >p4 j,k.tosetuptwo tunnes, assigning channes 1, 3} to one tunne and channes 2, 4} to the other tunne achieves the axiu expectation of successfu transission on path section i, j, k}. Proof: Let the success probabiities on the channes be p 1 i,j =1 p1 i,j, p2 i,j =1 p2 i,j, p3 j,k =1 p3 j,k,and p4 j,k = 1 p 4 j,k.wehave p1 i,j < p2 i,j and p3 j,k < p4 j,k. Coparing the success probabiities of the channe assignent given in Theore 1 and that of the aternative assignent, we have p 1 i,j p3 j,k + p2 i,j p4 j,k p1 i,j p4 j,k p2 i,j p3 j,k =( p1 i,j p2 i,j )( p3 j,k p 4 j,k ) > 0. The resut foows. According to Theore 1, a greedy approach, which aways chooses the channe with the owest oss rate at each ink when setting up tunnes aong a path, produces the optia overa success probabiity. More specificay, when there is ony one tunne to be set up aong a path, the tunne shoud consist of the ost reiabe channes avaiabe at each ink aong the path. When there are utipe tunnes to set up aong a path, tunne 1 shoud consist of the ost reiabe channes that are avaiabe at each ink; tunne 2 shoud consist of the second ost reiabe inks avaiabe at each ink; and so forth. Define the set of oss rates of the avaiabe channes on ink i, j} as Λ i,j = p i,j Ω i,j}. The greedy agorith is given in Tabe II, with which each video source node soves Probe OPT-CS for each feasibe path. Lines 2 3 in Tabe II checks if there is ore channes to assign and the agorith terinates if no channe is eft. In Lines 4 11, inks with ony one avaiabe channe are assigned to tunne r and the neighboring inks with the sae avaiabe channes are reoved due to constraint (21). In Lines 12 18, inks with ore than two channes are grouped to be assigned ater. In Lines 19 21, the avaiabe channe with the owest packet oss rate is assigned to tunne r at each unaocated ink, according to Theore 1. To avoid co-channe interference, the sae channe on neighboring inks is reoved as in Lines C. Sove the Path Seection Subprobe To sove Probe OPT-PS, we first reax binary variabes y h to aow the take rea vaues in [0,1] and obtain the foowing

7 HU and MAO: STREAMING SCALABLE VIDEOS OVER MULTI-HOP COGNITIVE RADIO NETWORKS 3507 TABLE III DISTRIBUTION ALGORITHM FOR PATH SELECTION 1: Initiaization: set τ =0, e g(0) > 0 andstepsizes [0, 1]; 2: Each source ocay soves the ower eve probe in (32); if (F h g dg,h eg(τ)) > 0) yh = y h + s, yh =iny h, 1}; ese y h = y h s, yh =axy h, 0}; 3: Broadcast soution y h(e(τ)); 4: Each source updates e according to (34) and broadcasts e(τ +1) through the coon contro channe; 5: τ τ+1 and go to Step 2 unti terination criterion is satisfied; reaxed path seection probe (OPT-rPS): axiize: subject to: in e 0 h F h f(y) = y h (30) h P w g,h yh 1, for a g 0 y h 1, for a h,. We then introduce positive Lagrange Mutipiers e g for the path seection constraints in Probe OPT-rPS and obtain the corresponding Lagrangian function: L(y, e) = h F h yh + g e g(1 h wg,h yh ) (31) = h (F h y h g wg,h yh e g )+ g e g = h Lh (y h, e)+ g e g. Probe (31) can be decouped since the doains of y h s do not overap. Reaxing the couping constraints, it can be decoposed into two eves. At the ower eve, we have the foowing subprobes, one for each path P h, ax L h 0 y h 1 (yh h, e) =F yh g wg,h yh e g. (32) At the higher eve, by updating the dua variabes e g, we can sove the reaxed dua probe: ( (y ) h ), e + g e g, (33) h Lh q(e) = where ( ) y h is the optia soution to (32). Since the soution to (32) is unique, the reaxed dua probe (33) can be soved using the foowing subgradient ethod that iterativey updates the Lagrange Mutipiers [16]: e g (τ +1)= [ e g (τ) α(τ)(1 ] + h wg,h yh ), (34) where τ is the iteration index, α(τ) is a sufficienty sa positive step size and [x] + denotes axx, 0}. The pseudo code for the distributed agorith is given in Tabe III. D. Optiaity and Convergence Anaysis The distributed agorith in Tabe III iterativey updates the dua variabes unti they converge to stabe vaues. In this section, we first prove that the soution obtained by the distributed agorith is aso optia for the origina path seection probe OPT-PS. We then derive the convergence condition for the distributed agorith. Fact 1 ([16]): Consider a inear probe invoving both equaity and inequaity constraints axiize: a x (35) subject to: h 1 x = b 1,, h x = b g 1x c 1,, g rx c r, where a, h i,andg j are coun vectors in R n, b i s and c j s are scaars, and a is the transpose of a. For any feasibe point x, the set of active inequaity constraints is denoted by A(x) = j g j x = c j }.Ifx is a axiizer of inequaity constrained probe (35), x is aso a axiizer of the foowing equaity constrained probe: axiize: a x (36) subject to: h 1x = b 1,, h x = b g j x = c j, j A(x). Lea 1: The optia soution for the reaxed pria probe OPT-rPS in (30) is aso feasibe and optia for the origina Probe OPT-PS in (29). Proof: According to Fact 1, the inearized probe of OPT-PS, i.e., OPT-rPS, can be rewritten as an equaity constrained probe in the foowing for: axiize: F y (37) subject to: w jy =1, j A(y ) (38) 0 y h 1, for a h,, where F, w j s, and y are coun vectors with eeents F h, w g,h,andyh, respectivey. We appy Gauss-Jordan eiination to the constraints in (38) to sove for y. Since there is not sufficient nuber of equations, soe y h s are free variabes (denoted as y f i ) and the rest are dependent variabes (denoted as yj d ). Assuing there are r free variabes, the dependent variabes can be written as inear cobinations of the free variabes after Gauss-Jordan eiination, as yj d = r i=1 wi jy f i + b j,j A(yi ). (39) Due to Gauss-Jordan eiination and binary vectors w j s, w j i and b j in (39) are a integers. Therefore, if a the free variabes y f i attain binary vaues, then a the dependent variabes yj d coputed using (39) wi aso be integers. Since 0 yj d 1, being integers eans that they are either 0 or 1, i.e., binaries. That is, such a soution wi be feasibe. Next we substitute (39) into probe (37) to eiinate a the dependent variabes. Then we obtain a unconstrained probe with ony r free variabes, as axiize: r F i=1 i y f i + b 0 (40) Since the free variabes y f i s take vaue in 0, 1}, this probe can be easiy soved as foows. If the coefficient F i > 0, we set y f i = 1;otherwise,if Fi < 0, wesety f i = 0. Thus (40) achieves its axiu objective vaue. Once a the free variabes are deterined with their optia binary vaues, we coputes the dependent variabes using (39), which are aso binary as discussed above. Thus we obtain a feasibe soution, which is optia. Lea 2: If the reaxed pria Probe OPT-rPS in (30) has an optia soution, then the reaxed dua probe (33) aso has an optia soution and the corresponding optia vaues of the two probes are identica. Proof: By definition, the probes in (31) and (33) are pria/dua probes. The pria probe aways has an optia soution because it is bounded. Since Probe OPTrPS is an LP probe, the reaxed dua probe is aso

8 3508 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 11, NOVEMBER 2010 bounded and feasibe. Therefore the reaxed dua probe aso has an optia soution. We have the strong duaity if the pria probe is convex, which is the case here since Probe OPT-rPS is an LP probe. We have Theore 2 on the optiaity of the path seection soution, which foows naturay fro Leas 1 and 2. Theore 2: The optia soution to the reaxed dua probe (32) and (33) is aso feasibe and optia to the origina path seection Probe OPT-PS given in (29). As discussed, the reaxed dua probe (33) can be soved using the subgradient ethod that iterativey updates the Lagrange Mutipiers. We have the foowing theore on the convergence of the distributed agorith given in Tabe III. Theore 3: Let e be the optia soution. The distributed agorith in Tabe III converges if the step sizes α(τ) in (34) satisfy the foowing condition: 0 <α(τ) < 2[q(e(τ)) q(e )] G(τ) 2, for a τ, (41) where G(τ) is the gradient of q(e(τ)). Proof: Since q(e(τ)) is a inear function, we have subgradient equaity, as q(e(τ)) q(e )=[e(τ) e ] G(τ). It then foows that e(τ) α(τ)g(τ) e 2 = e(τ) e 2 2α(τ)[e(τ) e ] G(τ)+(α(τ)) 2 G(τ) 2 = e(τ) e 2 2α(τ)[q(e(τ)) q(e )] + (α(τ)) 2 G(τ) 2. (42) If α(τ) satisfy (41), the su of the ast two ters in (42) is negative. It foows that, e(τ) α(τ)g(τ) e < e(τ) e. Since the projection operation is nonexpansive, wehave, e(τ +1) e = [e(τ) α(τ)g(τ)] + [e ] + e(τ) α(τ)g(τ) e < e(τ) e, which states the conditiona convergence of the agorith. Since the optia soution e is not known a priori, we use the foowing approxiation in the agorith: α(τ) = q(e(τ )) ˆq(τ ) G(τ ) 2,whereˆq(τ) is the current estiate for q(e ).We choose the ean of the objective vaues of the reaxed pria and dua probes for ˆq(τ). E. Practica Considerations The distributed agoriths are based on the fact that the coputation is distributed on each feasibe path. The OPT- CS agorith requires inforation on channe avaiabiity and packet oss rates at the inks of feasibe paths. The OPT-PS agorith coputes the pria variabe y h for each path and broadcasts Lagrangian utipiers over the contro channe to a the source nodes. We assue a perfect contro channe such that channe inforation can be effectivey distributed and shared, which is not confined by the tie sot structure [10]. We assue reativey arge tiescaes for the priary network tie sots, and sa to ediu diaeter for the CR network, such that there is sufficient tie for tiey feedback of channe inforation to the video source nodes and for s1 s2 d1 Priary Network 1 s3 Priary Network 3 d3 d2 Priary Network 2 Fig. 3. Topoogy of the uti-hop CR network. Note that ony video source nodes, video destination nodes, and those nodes aong the precoputed paths are shown in the topoogy. the distributed agoriths to converge. Otherwise, channe inforation can be estiated using (5) based on deayed feedback, eading to suboptia soutions. If the tie sot is too short, the distributed agorith ay not converge to the optia soution (see Fig. 7). We focus on deveoping the CR video fraework in this paper, and wi investigate these issues in our future work. V. SIMULATION STUDIES A. Methodoogy and Siuation Settings We ipeent the proposed agoriths with a cobination of C and MATLAB (i.e., for soving the reaxed NLP probes), and evauate their perforance with siuations. For the resuts reported in this section, we have K =3priary networks and M =10channes. There are 56, 55, and 62 CR users in the coverage areas of priary networks 1, 2, and 3, respectivey. The U 1 sare[ ](i.e., five users sense channe 1, four users sense channe 2, and so forth); the U 2 sare[ ],andthe U 3 s are [ ]. The topoogy is shown in Fig. 3. We choose L p = 100, T s = 0.02 and N G = 10.The channe utiization is η k = 0.6 for a the channes. The probabiity of fase aar is ε k =0.3 and the probabiity of iss detection is δ k =0.2 for a and k, uness otherwise specified. Channe paraeters λ k and μ k are set between (0, 1). The axiu aowed coision probabiity γ k is set to 0.2 for a the M channes in the three priary networks. We consider three video sessions, each streaing a video in the Coon Interediate Forat (CIF, ), i.e., Bus to destination 1, Forean to destination 2, and Mother & Daughter to destination 3. The frae rate is 30 fps, and a GOP consists of 10 fraes. The duration of a tie sot is 0.02 seconds and each GOP shoud be deivered in 0.2 seconds (i.e., 10 tie sots). We copare four schees in the siuations: (i) the upperbounding soution by soving the reaxed version of Probe OPT-CRV using an NLP sover, (ii) the proposed distributed agorith in Tabes II and III, (iii) the sequentia fixing agorith given in Tabe I, which coputes a ower-bounding soution, and (iv) a greedy heuristic where at each hop, the ink with the ost avaiabe channes is used. Each point in the figures is the average of 10 siuation runs, with 95% confidence intervas potted as error bars in the figures. The 95% confidence intervas are negigibe in a the figures.

9 HU and MAO: STREAMING SCALABLE VIDEOS OVER MULTI-HOP COGNITIVE RADIO NETWORKS e 4 (τ) Upper Bound Distributed Sequentia Fixing Heuristic Schee Lagrangian Mutipier e e 3 (τ) e 2 (τ) Contro overhead Contro Overhead PSNR (db) Fig e 0 1 (τ) Iteration index (τ) Iustrate the convergence of the distributed agorith. B. Siuation Resuts 1) Agorith Perforance: To deonstrate the convergence of the distributed agorith, we pot the traces of the four Lagrangian utipiers in Fig. 4. We observe that a the Lagrangian utipiers converge to their optia vaues after 76 iterations. We aso pot the contro overhead as easured by the nuber of distinct broadcast essages for e i (τ) using the y-axis on the right-hand side. The overhead curve increases ineary with the nuber of iterations and gets fat (i.e., no ore broadcast essage) when a the Lagrangian utipiers converge to their optia vaues. We exaine the ipact of spectru sensing errors in Fig. 5. We test six sensing error cobinations ε,δ } as foows: 0.1, 0.5}, 0.2, 0.3}, 0.3, 0.2}, 0.5, 0.11}, 0.7, 0.06}, and 0.9, 0.02}, and pot the average PSNR vaues of the Forean session. It is interesting to see that the best video quaity is achieved when the fase aar probabiity ε is between 0.2 and 0.3. Since the two error probabiities are correated, increasing one wi generay decrease the other. With a arger ε, CR users are ore ikey to waste spectru opportunities that are actuay avaiabe, eading to ower bandwidth for videos and poorer video quaity, as shown in Fig. 5. On the other hand, a arger δ ipies ore aggressive spectru access and ore severe interference to priary users. Therefore when ε is ower than 0.2 (and δ is higher than 0.3), the CR nodes theseves aso suffer fro the coisions and the video quaity degrades. 2) Ipact of Priary Network Paraeters: In Fig. 6, we exaine the ipact of channe utiization η on received video quaity. We focus on Session 2 with the Forean sequence. The average PSNRs achieved by the four schees are potted when η is increased fro 0.6 to 0.9 for a icensed channes. Intuitivey, a saer η aows ore transission opportunities for CR nodes, eading to iproved video quaity. This is iustrated in the figure where a the four curves decrease as η gets arger. The distributed schee achieves PSNRs very cose to that obtained by sequentia fixing, and both of the are cose to the upper bound. The heuristic schee is inefficient in expoiting the avaiabe spectru even when the channe utiization is ow. 50 Fig. 5. PSNR (db) Fase Aar Probabiity (ε) Fig. 6. Video PSNR versus spectru sensing error. Upper Bound Disbtributed Sequentia Fixing Heuristic Schee Channe Utiization (η) Video PSNR versus priary user channe utiization η. As discussed, the tie sot duration is aso an iportant paraeter that ay affect the convergence of the distributed agorith. In Fig. 7, we keep the sae network and video session settings, whie increasing the tie sot duration as 4 s, 10 s, 20 s, 40s and 100 s. For a given tie sot duration, we et the distributed agorith run for 5% of the tie sot duration, starting fro the beginning of the tie sot, and then stop. The soution that the agorith produces when it is stopped wi be used for video transission in the reainder of this tie sot. It can be seen that when the tie sot is 4 s, the agorith does not converge after 5% 4=0.2 s, and the PSNR produced by the distributed agorith is ow (but sti higher than that of the heuristic agorith). When the tie sot duration is sufficienty arge (e.g., over 10 s), the agorith can converge and the proposed agorith produces very good video quaity as copared to the upper bound and the ower bound given by the sequentia fixing agorith. 3) Coparison of MPEG-4 FGS and H.264/SVC MGS Videos: Finay, we copare MPEG-4 FGS and H.264/SVC MGS videos, whie keeping the sae settings. It has been shown that H.264/SVC has better rate-distortion perforance than MPEG-4 FGS due to the use of efficient hierarchica prediction structures, the inter-ayer prediction echaniss, iproved drift contro echanis, and the efficient coding schee in H.264/AVC [5]. Athough MGS has Network Abstraction Layer (NAL) unit-based granuarity, it achieves siiar rate-distortion perforance as H.264/SVC FGS [5].

10 3510 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 9, NO. 11, NOVEMBER Upper Bound: MGS Distributed: MGS PSNR (db) Upper Bound Distributed Sequentia Fixing Heuristic Schee PSNR (db) Upper Bound: FGS Distributed: FGS Fig. 7. PSNR (db) Tie Sot Duration (s) Ipact of tie sot duration on received video quaity. Upper bound: MGS Distributed: MGS Upper bound: FGS Distributed: FGS Channe Utiization (η) Fig. 8. Coparison of MPEG-4 FGS video with H.264/SVC MGS video under various channe utiizations. We pot the upper bounds and the distributed agorith resuts in Figs. 8 and 9 for various channe utiizations and fase aar probabiities, respectivey. Fro the figures, it can be observed that there is a gap about 2.5 db between the H.264/SVC MGS and MPEG-4 FGS curves, which ceary deonstrates the rate-distortion efficiency of MGS over MPEG-4 FGS. The proposed agorith can hande both MGS and FGS videos, and the sae trend is observed in both cases. VI. RELATED WORK The high potentia of CRs has attracted considerabe interest fro the wireess counity [1], [2], [17]. The ainstrea CR research has been focused on spectru sensing and dynaic spectru access issues [9] [11], [13], [18]. Severa papers have addressed the ipact of spectru sensing errors on the design of spectru access schees [18] [21]. The approach of iterativey sensing a seected subset of avaiabe channes has been adopted in the design of CR MAC protocos (e.g., see [10], [13], [18]). Our work is copeentary to this cass of work by providing an iportant appication for the enhanced spectru efficiency achieved by spectru sensing and access schees reported in the iterature Fase Aar Probabiity (ε) Fig. 9. Coparison of MPEG-4 FGS video with H.264/SVC MGS video under various fase aar probabiities. Muti-hop SDR or CR networks have been studied in a few recent works [15], [22], [23]. The authors foruate crossayer optiization probes considering factors fro the PHY up to the transport ayer. Distributed agoriths are deveoped by appying the dua decoposition technique [16], [24]. We adopt siiar ethodoogy in this paper, but address the ore chaenging probe of streaing rea-tie videos. The probe of QoS provisioning in CR networks has been considered in a few papers [10], [20], [25], where the focus is sti on the so-caed network-centric etrics such as axiu throughput and deay [10], [20]. In a recent work [20], Urgaonkar and Neey derive an interesting deay throughput trade-off for a uti-ce cognitive radio network, whie priary user protection is achieved by stabiizing a virtua coision queue. In [25], a gae-theoretic fraework is described for resource aocation for utiedia transissions in spectru agie wireess networks. In this interesting work, each wireess station pays a resource anageent gae, which is coordinated by a network oderator. A echanis-based resource anageent schee deterines the aount of transission tie to be aocated to various users on different frequency bands such that certain goba syste etrics are optiized. In our prior work [3], we consider video uticast in an infrastructure-based CR network. We present effective greedy heuristic agoriths for scheduing video data, with proved optiaity bound and ow coputationa copexity. In this paper, we consider the chaenging case of uti-hop CR networks, where distributed agoriths are highy appeaing. VII. CONCLUSION We studied the probe of streaing utipe scaabe videos in a uti-hop CR network. The probe foruation considered spectru sensing and sensing errors, spectru access and priary user protection, video quaity and fairness, and channe/path seection for concurrent video sessions. We first soved the foruated MINLP probe using a sequentia fixing schee that produces ower and upper bounds on the achievabe video quaity. We then appied dua decoposition

11 1 HU and MAO: STREAMING SCALABLE VIDEOS OVER MULTI-HOP COGNITIVE RADIO NETWORKS 3511 to derive a distributed agorith, and anayzed its optiaity and convergence perforance. Our siuations vaidated the efficacy of the proposed schee. REFERENCES [1] I. Akyidiz, W. Lee, M. Vuran, and S. Mohanty, NeXt generation/dynaic spectru access/cognitive radio wireess networks: a survey," Coputer Netw. J., vo. 50, no. 9, pp , Sep [2] Q. Zhao and B. Sader, A survey of dynaic spectru access," IEEE Signa Process. Mag., vo. 24, no. 3, pp , May [3] D. Hu, S. Mao, and J. Reed, On video uticast in cognitive radio networks," in Proc. IEEE INFOCOM 09, Rio de Janeiro, Brazi, Apr. 2009, pp [4] H. Radha, M. van der Schaar, and Y. Chen, The MPEG-4 fine-grained scaabe video coding ethod for utiedia streaing over IP," IEEE Trans. Mutiedia, vo. 3, no. 1, pp , Mar [5] M. Wien, H. Schwarz, and T. Oebau, Perforance anaysis of SVC," IEEE Trans. Circuits Syst. Video Techno., vo. 17, no. 9, pp , Sep [6] N. Lanean, D. Tse, and G. Worne, Cooperative diversity in wireess networks: efficient protocos and outage behavior," IEEE Trans. Inf. Theory, vo. 50, no. 11, pp , Nov [7] R. Kesavan and D. K. Panda, Efficient uticast on irreguar switchbased cut-through networks with up-down routing," IEEE Trans. Parae Distrib. Syst., vo. 12, no. 8, pp , Aug [8] R. Raanathan, Chaenges: a radicay new architecture for next generation obie ad-hoc networks," in Proc. ACM MobiCo 05, Coogne, Gerany, Sep. 2005, pp [9] J. Jia, Q. Zhang, and X. Shen, HC-MAC: a hardware-constrained cognitive MAC for efficient spectru anageent," IEEE J. Se. Areas Coun., vo. 26, no. 1, pp , Jan [10] H. Su and X. Zhang, Cross-ayer based opportunistic MAC protocos for QoS provisionings over cognitive radio wireess networks," IEEE J. Se. Areas Coun., vo. 26, no. 1, pp , Jan [11] A. Motaedi and A. Bahai, MAC protoco design for spectruagie wireess networks: stochastic contro approach," in Proc. IEEE DySPAN 07, Dubin, Ireand, Apr. 2007, pp [12] H. Mahoud, T. Yücek, and H. Arsan, OFDM for cognitive radio: erits and chaenges," IEEE Wireess Coun., vo. 16, no. 2, pp. 6-14, Apr [13] Q. Zhao, S. Geirhofer, L. Tong, and B. Sader, Opportunistic spectru access via periodic channe sensing," IEEE Trans. Signa Process., vo. 36, no. 2, pp , Feb [14] M. van der Schaar, S. Krishnaachari, S. Choi, and X. Xu, Adaptive cross-ayer protection strategies for robust scaabe video transission over WLANs," IEEE J. Se. Areas Coun., vo. 21, no. 10, pp , Dec [15] Y. Hou, Y. Shi, and H. Sherai, Spectru sharing for uti-hop networking with cognitive radios," IEEE J. Se. Areas Coun., vo. 26, no. 1, pp , Jan [16] D. P. Bertsekas, Noninear Prograing. Athena Scientific, [17] Y. Zhao, S. Mao, J. Nee, and J. H. Reed, Perforance evauation of cognitive radios: etrics, utiity functions, and ethodoogies," Proc. IEEE, vo. 97, no. 4, pp , Apr [18] D. Hu and S. Mao, Design and anaysis of a sensing error-aware MAC protoco for cognitive radio networks," in Proc. IEEE GLOBECOM 09, Honouu, HI, Nov./Dec [19] Y. Chen, Q. Zhao, and A. Swai, Joint design and separation principe for opportunistic spectru access in the presence of sensing errors," IEEE Trans. Inf. Theory, vo. 54, no. 5, pp , May [20] R. Urgaonkar and M. Neey, Opportunistic scheduing with reiabiity guarantees in cognitive radio networks," IEEE Trans. Mobie Coput., vo. 8, no. 6, pp , June [21] T. Shu and M. Krunz, Throughput-efficient sequentia channe sensing and probing in cognitive radio networks under sensing errors," in Proc. ACM MobiCo 09, Beijing, China, Sep. 2009, pp [22] Y. Hou, Y. Shi, and H. Sherai, Optia spectru sharing for utihop software defined radio networks," in Proc. IEEE INFOCOM 07, Anchorage, AK, Apr. 2007, pp [23] Z. Feng and Y. Yang, Joint transport, routing and spectru sharing optiization for wireess networks with frequency-agie radios," in Proc. IEEE INFOCOM 09, Rio de Janeiro, Brazi, Apr. 2009, pp [24] D. Paoar and M. Chiang, A tutoria on decoposition ethods for network utiity axiization," IEEE J. Se. Areas Coun., vo. 24, no. 8, pp , Aug [25] A. Fattahi, F. Fu, M. van der Schaar, and F. Paganni, Mechanisbased resource aocation for utiedia transission over spectru agie wireess networks," IEEE J. Se. Areas Coun., vo. 25, no. 3, pp , Apr Dongin Hu received the M.S. degree fro Tsinghua University, Beijing, China, in 2007 and the B.S. degree fro Nanjing University of Posts and Teecounications, Nanjing, China in 2004, respectivey, a in eectrica engineering. Since 2007, he has been pursuing a Ph.D. degree in the Departent of Eectrica and Coputer Engineering, Auburn University, Auburn, AL. His research interests incude cognitive radio networks, cross-ayer optiization, agorith design for wireess network and utiedia counications. Shiwen Mao (S 99-M 04-SM 09) received a Ph.D. in eectrica and coputer engineering fro Poytechnic Institute of New York University, Brookyn, NY (forery known as Poytechnic University) in He was a research staff eber with IBM China Research Lab fro 1997 to He was a Research Scientist in the Bradey Departent of Eectrica and Coputer Engineering at Virginia Poytechnic Institute and State University (Virginia Tech), Backsburg, VA fro 2003 to Currenty, he is an Assistant Professor in the Departent of Eectrica and Coputer Engineering, Auburn University, Auburn, AL. His research interests incude cross-ayer optiization of wireess networks and utiedia counications, with current focus on cognitive radio networks and free space optica networks. He is on the Editoria Board of IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, IEEE COMMUNICA- TIONS SURVEYS AND TUTORIALS, Esevier Ad Hoc Networks Journa, Wiey Internationa Journa of Counication Systes, andicst Transactions on Mobie Counications and Appications. He chairs the Interest Group on Cross-ayer Design for Mutiedia Counications of IEEE Counications Society s Mutiedia Counications Technica Coittee. Dr. Mao is a coauthor of TCP/IP Essentias: A Lab-Based Approach (Cabridge, U.K.: Cabridge University Press, 2004). Dr. Mao received the US Nationa Science Foundation (NSF) Facuty Eary Career Deveopent Award (CAREER) in 2010, the 2004 IEEE Counications Society Leonard G. Abraha Prize in the Fied of Counications Systes, and the Best Paper Runner-up Award fro the Fifth Internationa Conference on Heterogeneous Networking for Quaity, Reiabiity, Security and Robustness (QShine) He hods one US patent.